Ian Coe from Tonic.ai: synthetic data, AI training, and enterprise fine-tuning
Six months ago, Tonic.ai customers said they would never fine-tune and would wait for foundation models to improve; now cheaper specialized models and uneven frontier capabilities are changing that calculation.
Ian Coe is the co-founder and CEO of Tonic.ai, which began by making safe, realistic data available for software development and testing and now also serves teams training AI models.
Recorded at Heavybit DevGuild, Ian explains how structured development data differs from the unstructured material used to train models, how synthetic datasets are shaped by people rather than conjured from nothing, and why healthcare and financial-services companies need de-identification and compliance work before proprietary data can become useful training material.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
How did Tonic expand from a developer tool into AI training data?
We started by making data available for technical workflows, initially testing and developing software. As AI created new data needs, the customer base expanded from developers to people training models. The common thread is making existing data usable or creating representative data when the original cannot safely be used.
How do software-development and model-training data differ?
For software development, the structure and relationships in a database are central because the goal is to reproduce staging and development environments. AI training more often involves unstructured text. A healthcare or financial-services company may have a corpus of support conversations containing sensitive information, so that material has to be de-identified before it can train a model.
Can Tonic create synthetic data from nothing?
Tonic Fabricate can generate a dataset from a description, small examples or a profile of an existing database. It is particularly useful for reinforcement-learning data. Models accelerate the process, but they are not autonomously training themselves: a human inspects and shapes the generated data toward the world they want to represent.
Can synthetic data perform as well as real data for reinforcement learning?
Our published work shows synthetic data performing as well as real data for some reinforcement-learning tasks. The setup still needs tasks, an environment and an evaluation. In one experiment involving email retrieval, the team used generated data to fine-tune an open model and achieved results that were better in a number of ways than the earlier real-data approach.
Why are enterprises reconsidering fine-tuning?
About six months ago, customers often said they would never fine-tune because foundation labs would eventually deliver the capability. The conversation has changed. Frontier performance is likely to remain jagged—excellent at coding, perhaps not sufficient for a specific insurance-claims task—and inference may become more expensive. An open model that trails the frontier generally can still be cheaper and better at the one specialized task after training on company data.
Where do Fabricate and Tonic Textual fit into that workflow?
Fabricate creates fully synthetic data when a company has no suitable corpus or needs to fill its gaps. Tonic Textual starts from unstructured source material and identifies or transforms sensitive information so the data can be used. Together they support teams whether they need to create a training environment from scratch or safely unlock data they already possess.
How does de-identification work in healthcare?
A healthcare organization can follow HIPAA safe harbor, which removes so much information that the data may lose much of its value, or use expert determination. In the latter approach, an expert concludes that the transformed dataset is statistically safe. Tonic performs the manipulations that let customers retain more utility while meeting that process.
What separates enterprises succeeding with AI from those getting stuck?
The ambition is widespread, and some customers have fine-tuned a model and put it into an operational business process. Others remain caught in compliance and data-preparation work. Access to models is only one part of adoption; organizations also need a lawful, technically useful path from their sensitive internal data to a production system.